Energy consumption is a key challenge in current artificial intelligence (AI) hardware. AI training and deploying large language models require vast amounts of memory and storage, and each GPU node in a training cluster can require a large amount of energy. AI training has become one of the most resource-intensive computing tasks in the modern era.
Training runs can consume over 27GWh of energy, with GPU clusters facing significant reliability challenges, including frequent component failures. To address that, you need devices with extremely low currents, stability and the ability to switch between distinct states. Training and deploying Large Language Models (LLMs) also demands immense infrastructure with specialized, high-bandwidth memory (HBM) to process massive datasets and tens of thousands of GPUs.
Current AI systems rely on conventional computer chips that shuttle data back and forth between memory and processing units. This constant movement consumes large amounts of electricity. This further creates heat and wastes power. Brain-inspired, or neuromorphic, computing is an alternative way to process information that could reduce energy use by as much as 70% by storing and processing information in the same place – and doing so with extremely low power.
Mimicking Synaptic Behavior
Our brains process and store information in the same place: the synapse. Similarly, memristors are two-terminal devices that can store and process data in the same physical location. Researchers at the University of Cambridge have developed a new type of hafnium oxide memristor that operates at switching currents much lower than conventional oxide-based devices. The researchers published a paper in the journal Science Advances titled “HfO2-Based Memristive Synapses with Asymmetrically Extended P-N Heterointerfaces for Highly Energy-Efficient Neuromorphic Hardware.” The paper details a nanoelectronic component designed to significantly reduce the energy consumption of AI hardware.
A memristor is a two-terminal passive circuit element. It retains memory of the amount of charge that has flowed through it. Its resistance changes based on the history of voltage and current. Memristors can be used in non-volatile memory storage. They are key components in neuromorphic computing, mimicking synaptic behavior. Memristors can enable faster and more efficient data processing.
Memristors can enable faster and more efficient data processing. The new device incorporates strontium and titanium to create internal p-n junctions that act as smooth electronic gates inside the oxide where the layers meet. Because our devices switch at the interface, they show outstanding uniformity from cycle to cycle and from device to device.
The Cambridge team used a self-assembled, multi-component p-type Hf(Sr,Ti)O2 thin film for energy-efficient, resistive switching–based neuromorphic type devices. These memristors are said to feature ultra-low switching currents and exceptional cycle-to-cycle and device-to-device uniformity and retention. Switching is achieved by adjusting an energy barrier at the interface, providing uniformity across cycles and devices. Because the devices switch at the interface, they show uniformity from cycle to cycle and from device to device.
This breakthrough reduces power consumption by using switching currents a million times smaller than those of older technology. In the future, this brain-mimicking chip could slash AI’s energy appetite significantly.
700°C Hurdle
Laboratory tests confirm these devices can endure tens of thousands of cycles while holding their data for about a day. However, there are still some challenges to overcome. The current fabrication process requires temperatures of around 700°C – higher than standard semiconductor manufacturing tolerances. That is far too hot for standard semiconductor manufacturing, which requires much cooler conditions to avoid melting delicate components.
Still, this materials-engineering strategy addresses energy consumption and variability in existing memristors, opening a pathway toward energy-efficient neuromorphic computing systems.
The research was supported in part by the Swedish Research Council (VR), the European Research Council, the Royal Academy of Engineering, the Royal Society and UK Research and Innovation (UKRI). A patent application has been filed by Cambridge Enterprise, the University’s innovation arm.
Follow TTI, Inc. on LinkedIn for more news and market insights.
Statements of fact and opinions expressed in posts by contributors are the responsibility of the authors alone and do not imply an opinion of the officers or the representatives of TTI, Inc. or the TTI Family of Specialists.
Follow TTI, Inc. - Europe on LinkedIn for more news and market insights.
Statements of fact and opinions expressed in posts by contributors are the responsibility of the authors alone and do not imply an opinion of the officers or the representatives of TTI, Inc. or the TTI Family of Specialists.
